Wearable Sensing and Physical Exercise Recognition

As society ages, it becomes critically important to maintain the physical and mental health of the elderly population. This paper aims at using wearable sensors to automatically recognize three physical exercises (squatting, jumping jacks, flicking back), which are known to assist the elderly with balance, strength and endurance. The system will automatically track and analyze whether the actions are meeting the exercise standards using the wearable sensor and machine learning based recognition techniques. Data were collected from four different positions (chest, wrist, waist, ankle) nine times for ten users with the 3-axis accelerometer sensor with 100 Hz sampling frequency. For the simplicity of testing, each exercise lasted 20 seconds. For the comparative analysis, four kinds of models (decision tree, KNN, SVM, Naïve Bayes) were used to obtain the results of recognition. The study shows that the recognition result is not only related to the location of the sensor position but also related to the classification model. In general, the Decision Tree model performed better than other models for all of the four locations, with an average accuracy of 92.52%.